For a decade, Digital Product Creation (DPC) promised a workflow that would flow smoothly from concept to product: one digital thread, tools that plugged into each other, no friction. But most of that promise never materialised.
Last year, H&M used a brand-level strategy shift, from retailer to fashion brand, as the reason to stop patching around that gap and start over. They pulled 3D out of product design entirely, built an internal AI design tool called StyleGen with four people in four months, and now run it across every H&M portfolio brand: 350+ monthly active users with roughly 100K pieces of content generated a month.
Bhargava Ram Kummamuru, Head of H&M's Digital Product Creation, and standout speaker at The Fashion Tech Show 2026 in both London and NYC, gives us an honest account of what worked, what didn't, and why "buy-versus-build" was the wrong question to be asking all along.
What follows is an editorial summary, key takeaways and the full event video from last month's The Fashion Tech Show NYC for anyone who missed out.
A decade of broken promises
H&M's transformation started with the question: has the industry actually simplified DPC, or just kept adding tools and calling the mess "innovation"? The last decade promised seamless 3D from concept to product, a unified digital twin, and tools that plugged into each other. But, according to Ram, almost none of it held up.
We made claims in the past ten years that we would have seamless 3D design from concept to product. That magically, when you do 3D, everything will flow fluently through the whole thing.
But what actually showed up instead was rising cost of ownership, constant tool-switching, and change management that never really ended. The industry, as he put it, kept "chopping processes into tools and hoping it magically works," patching gaps with automation instead of challenging whether the approach, as a whole, was fundamentally wrong.
Strategy before technology
Technology wasn't the trigger, strategy was. When H&M brought in new leadership in 2024 and 2025 and repositioned itself from retailer to fashion brand, design's role shifted with it, away from owning every detail of a product in-house, towards a more collaborative process worked out with suppliers.
Ram has a model for how he approaches change like this: start with purpose, the reason you're doing this at all, then process, and only at the end, tools and roles. Most organisations do it backwards, reaching for tools first and hoping the purpose sorts itself out later.
The same instinct shapes how he decides what's worth changing at all. He sorts everything into a problem statement or an opportunity statement, and most transformation conversations start from a problem someone's already flagged. AI wasn't one of those.
Nobody has really said that's a problem. We saw the market evolving, something shifting and moving in that direction. And we said, let's take this opportunity and see if it actually fits for us.
Pulling 3D out of product design
That framework is what gave Ram's team room to make a decision that would've been hard to justify otherwise: pulling 3D out of product designers' hands entirely. The reasoning was concrete, not theoretical. A designer had told him she sometimes avoided complex product ideas simply because they were hard to execute in 3D.
I hated that because limitations of a tool shouldn't be limiting creative intent. That should not be the outcome of tools. Tools are supposed to enable, not constrain.
Building StyleGen
What replaced 3D was StyleGen, an internal AI design tool that lets designers generate garment concepts from a text prompt, iterate on fabrics and details, produce technical sketches for suppliers, combine hand-drawn prints into repeat patterns, apply those prints to a garment, and visualise the finished piece on a model, all inside one workflow instead of bouncing between Photoshop, Illustrator, and 3D tools.
Four people built it in four months, working off a short list of principles and a simple mandate:
Keep it simple. Let's not build another thing that has millions of buttons, millions of operations, millions of workflows.
The team also drew a hard line between subjective work, like creative exploration, and objective work, like exact colour-matching, and chose to scale investment only as value got proven, rather than betting everything on one big build.
On the buy-versus-build question that most DPC conversations get stuck on, Ram was clear:
That is not the world we are living in anymore. We have to live in a world which is a combination of buy and build.
StyleGen shipped without H&M's custom colour-code system at first. That feature got added later, once usage data showed people were still bouncing out to Photoshop for it. The lesson: watch what people actually do with the tool, then build the next thing, instead of guessing at every need upfront.
Under the hood
The architecture is deliberately layered. At the front end sits a user interface built around natural language. Underneath that is an agent orchestration layer, which decides what kind of job a request actually is.
If the user says I want to generate an AI design, it uses a [generation] model. But if the user says I want to remove the background, that's not an AI job. That is just a simple automation job. Our agent understands to run that specific job.
Below the orchestration layer sit workflow services, which encode the specific processes designers actually need: generating colourways, image scaling, chat history, H&M's own colourway codes, in-painting, text-to-image and image-to-image generation, and background removal. And Ram gave a concrete example of why this distinction matters: ask for five different design options and the system treats each as a separate idea; ask for three colourways of the same design and it understands that's a single, deterministic task, not five random variations.
Each workflow calls on a layer of tools underneath it, from image generation models to H&M's colourway data to background-removal utilities and select third-party APIs. The whole thing runs on an Azure enterprise back end, with storage and analytics layered in so the team can actually measure performance and usage over time, not just guess at it.
Results that compound
The results look less like a flashy demo and more like steady, compounding adoption: 560+ users, 350+ monthly actives, roughly 700K pieces of content generated since launch, now running around 100K a month, across every product category and portfolio brand.
There's a small sign on floor two of an H&M store crediting a Weekday piece as the company's first AI-designed product to reach market. Ram brought it up as proof this went well past pilot stage.
Training the process, not the tool
The more interesting story is how adoption actually happened. Early on, H&M trained people on the tool itself: buttons, features, workflows. It didn't stick. Usage only took off once they switched to training people on process instead, bringing physical samples into the room, working iPad-in-hand through real design problems together.
Some of the best use cases came from the design community, not the product team. A materials development group started using the tool to pre-visualise fabric finishes for suppliers, and cut their sample-box iterations from six down to one. The Kids team put an AI-generated render next to the physical sample that came back from the factory and found the two close enough that the only real gap was photography lighting, not design accuracy.
AI is not a leveller
AI isn't an automatic leveller, either. H&M tested the same tool with engineers and with designers, and the engineers' output was noticeably worse.
Engineers made horrible designs. They were not at all good. And that is a fact, because engineers are not fashion designers.
The tool amplifies whatever judgement is already there. It doesn't replace it. Worth remembering before anyone assumes broad AI access is itself the win.
Staying clear of the old trap
Ram was also careful to avoid the trap that swallowed 3D a decade earlier: chasing flashy visuals or e-commerce-ready output while skipping past the harder problem of manufacturability.
The problem is not about taking to e-commerce or marketing, making a cool video, a cool collection.
The real work is translating a good image into something a factory can actually produce: patterns, measurement charts, fit.
H&M doesn't market any of its products as "AI-designed," for the same reason it never marketed products as "3D-designed." It's a tool, not a headline. The creative direction is still human, and a lot of what's on shelves right now used the tool without customers ever knowing.
On a separate leadership panel, Ram put the distinction plainly: "We don't do AI design. We use AI in design."
The shift inside engineering
The transformation reached inward too. Watching AI reshape design work pushed Ram's team to rethink how they build features internally. Subject matter experts on his team learned to vibe code working prototypes so they could show engineers exactly what they meant, instead of describing it and hoping.
That closed a familiar loop, build it, hear "that's not what I meant," rebuild, and cut feature lead time roughly in half in some cases.
What's next for DPC
H&M's ambition now is a single connected experience layer spanning ideation through tech-pack publishing, with the brand, not vendor tooling limits, setting the terms of what's possible.
Vendors cannot say you cannot do this. Then, I'm sorry, but we'll find somebody else who will do it. We'll not be stuck with you.
It's also why H&M runs vendor summits: bringing factories together to share direction and insight, without dictating which tool anyone should use.
He named three things the team now treats as non-negotiable: creativity and craftsmanship don't get compromised for tooling convenience, design has to move between steps without manual rework, and every investment gets judged on total cost of ownership across the whole DPC stack, not tool by tool.
Strip away the specific tools and one line holds the whole talk together: never let a tool decide what creative work is allowed to look like. It's a smaller claim than "AI transformation," but a much harder one to fake.
Key Takeaways
➤ H&M removed 3D entirely from product designers, not because 3D failed, but because it was quietly constraining creative decisions.
➤ The change was strategy-led, not tool-led. Ram's model: purpose, then process, then tools and roles, in that order.
➤ StyleGen was built by 4 people in 4 months. It now has 350+ monthly active users and generates roughly 100K pieces of content a month, across every portfolio brand.
➤ AI amplifies existing skill rather than levelling it. Same tool, noticeably worse output from engineers than from designers.
➤ Adoption came from switching "tool training" for "process training," and from use cases the design community found on its own.
➤ H&M doesn't market products as "AI-designed." It's a tool, not a selling point.
➤ The hard problem is still manufacturability, not generating a nice image. That's a lesson learned from 3D's decade-long shortfall.
➤ Non-negotiables going forward: creativity over tooling convenience, design fluidity, and total cost of ownership across the full stack.
➤ Ram sorts change into problem statements vs. opportunity statements. AI at H&M started as the latter, not a response to a stated pain point.
Watch the full talk
Jump to:
- 0:00 - Intro & Ram's role at H&M
- 3:45 - The industry's broken promises
- 8:55 - H&M's 2025 strategic shift & the 3 key decisions
- 15:05 - Building StyleGen: principles, live demo, H&M colour codes
- 21:24 - Real examples from H&M Home, Beauty, Kids, Weekday
- 28:34 - Agent orchestration, workflow services, tools, Azure backend
- 30:47 - Results, adoption & the tool- to process-training pivot
- 36:00 - SMEs "vibe coding" prototypes internally
- 40:10 - Where DPC is headed next
The Fashion Tech Show NYC will return in 2027. Register your interest here.
